Russ MillerCenter for Computational Research

Computer Science & Engineering SUNY-Buffalo

Hauptman-Woodward Medical Inst

IDF: Multi-Core Processing for HPC March 2005

University at Buffalo

The State University of New York

University at Buffalo The State University of New York CCRCenter for Computational Research

21st Century University

Embrace digital data-driven society Empower students to compete in knowledge-based economy Support research, scholarship, education, and community

outreach Deliver high-end cyberinfrastructure to enable efficient

Collection of data Management/Organization of data Analysis of data Visualization of data

University at Buffalo The State University of New York CCRCenter for Computational Research

Center for Computational Research 1999-2005 Snapshot

High-End Computing, Storage, Networking, and Visualization ~100 Research Groups in 37 Depts

Physical SciencesLife SciencesEngineeringScientific Visualization, Medical Imaging, Virtual Reality

13 Local Companies 10 Local Institutions

External Funding: $300M+ Total Leveraged WNY: $0.5B Deliverables

1100+ Publications Software, Media, Algorithms, Consulting,

Training, CPU Cycles…

University at Buffalo The State University of New York CCRCenter for Computational Research

SGI Altix3700 (0.4TF) 64 Processors (1.3GHz ITF2) 256 GB RAM 2.5 TB Disk

Apex Bioinformatics System Sun V880 (3), Sun 6800 Sun 280R (2) Intel PIIIs Sun 3960: 7 TB Disk Storage

HP/Compaq SAN 75 TB Disk; 190 TB Tape 64 Alpha Processors (400 MHz) 32 GB RAM; 400 GB Disk

Major Compute/Storage Resources

Dell Linux Cluster (2.9TF) 600 P4 Processors (2.4 GHz) 600 GB RAM; 40 TB Disk; Myrinet

Dell Linux Cluster (6TF) 4036 Processors (PIII 1.2 GHz) 2TB RAM; 160TB Disk; 16TB SAN

IBM BladeCenter Cluster (3TF) 532 P4 Processors (2.8 GHz) 5TB SAN

SGI Intel Linux Cluster (0.1TF) 150 PIII Processors (1 GHz) Myrinet

RFP (10-15TF) Pentium-Based Fast Interconnect Efficient Storage Management

University at Buffalo The State University of New York CCRCenter for Computational Research

CCR Visualization Resources

Fakespace ImmersaDesk R2 Portable 3D Device Onyx2: 6 R10000 @ 250MHz 2 IR2 Pipes; 3 64MB texture memory mgrs.

Tiled-Display Wall 20 NEC projectors: 15.7M pixels Screen is 11’7’ Dell PCs with Myrinet2000

Access Grid Nodes (2) Group-to-Group Communication Commodity components

SGI Reality Center 3300W Dual Barco’s on 8’4’ screen Onyx300: 10 R14000 @ 500MHz 2 IR4 Pipes; 1 GB texture mem per pipe

University at Buffalo The State University of New York CCRCenter for Computational Research

Multi-Core Applications Application Performance is Key

“Moore’s Law of Applications” Processor (Socket/Core) Performance Irrelevant Balanced System is Critical

Memory and Cache Access including SMPSystem I/ONo Choke Point

Multiprocessing is CriticalHuman Ingenuity Required/Available to Decompose AlgorithmsRededication to the Shared Memory Programming Problem

Multithreading is ImportantDifficult Problem in Scientific ComputingTools Required for Multi-Core/Multithreading Environment

University at Buffalo The State University of New York CCRCenter for Computational Research

Intel Multi-Core Systems Pentium Dual Core – 2005

Desktop Environment; Serves as Development Platform

Itanium Dual Core (Montecito) – 2005 Multithreading Dual Socket System = 8 Processors to OS 12MB L3 Cache Healthy Thermals

Xeon Dual Core (Dempsey) – 2006 Blackford/Greenfield Chipset – Balanced System

Dual Independent FSBFB-DIMMs More Efficient for Large Memory Configurations

Tukwila – 2007/8 More than 2 cores Common Chipset and Commodity Components (Memory, Power

Supplies, etc) Aimed at Consolidating Xeon and IPF

University at Buffalo The State University of New York CCRCenter for Computational Research

Benefits of Intel Multi-Core Systems

Application-Based Moore’s Law requiresParallel (Distributed and/or Shared Memory) andMultithreaded Implementations of Algorithms

Tools for Multithreading are CriticalHand-Coded Libraries (BLAS, SHMEM, FFT)Posix ThreadsCompiler Directives (OpenMP)Hybrid MPI-OpenMP

Balanced Systems are CriticalWell-funded applications will reap the benefit

University at Buffalo The State University of New York CCRCenter for Computational Research

Multi-Core Application Development

More Emphasis on Computational Science & Engineering at all Levels

Black Box Approach is Counter Productive Programming Skills Must be Improved Training Required at all Levels

High SchoolUndergraduateGraduatePost-Doctoral

Better & More Affordable Tools are Required Intel Provides Compilers and Libraries Free of Charge

(Academic Use/No Support) Intel is Major Player in Development Software (Costly)

University at Buffalo The State University of New York CCRCenter for Computational Research

www.ccr.buffalo.edu